关于「从最后一天往回推:贝尔曼逆推、网格检查与硬预算」,《新品冷启动的广告脉冲与口碑接棒》给出的定量结论是什么?
结论:贝尔曼逆推不假设任何形状,在允许中间金额的条件下独立得到同一条六天脉冲,最优价值 531904 与路径利润 531909 只差插值误差;从一百零一个到两千零一个网格点、从三档到二十一档,结论不变。决策表还可以读成一条随时间上升的阈值规则,用来应对销售偏离计划。硬预算下穷举所有摆法,最优永远是从第一天起连续打满。到此为止的结论都建立在一个关键假设上:广告效果对花费是线性的。如果头一千元最有效、此后边际递减,或者花得太少根本不起作用,最优路径还是一刀切吗?下一章回答。
数据口径:本书用巴斯扩散加广告控制的最优控制模型检验这一做法:广告以加法、线性方式作用于创新项并有每日上限时,庞特里亚金极大值原理与贝尔曼动态规划给出同一个答案,开头打满上限、在开关函数变负的那天一次停投,把接力棒交给口碑;示意参数下同额预算前六天打满比六十天均匀多赚 13.4 %。 书中的市场、毛利、扩散系数、广告效果、上限与产能均为示意参数;全部计算由书中程序生成;本书是非个人化经营建议,紧迫感只来自真实库存与真实截止。
对照数据表
下表数字由本章 Python 程序原样打印,可重新运行复现。 程序只用 Python 标准库与结论吻合的打印值: 531904, 531909
| DP value at x=0: 531904 | simulated path profit: 531909 |
| spend path days 0-9: 5000 5000 5000 5000 5000 5000 0 0 0 0 | |
| days at the cap: 6, days at a middle level: 0, total 30000 | |
| day | threshold x*_t (spend the cap while x is below it) |
| 0 | 0.108 |
| 2 | 0.118 |
| 4 | 0.127 |
| 6 | 0.137 |
| 8 | 0.148 |
| 10 | 0.160 |
| 20 | 0.228 |
| 40 | 0.408 |
核心公式
\[W_t(x)=\max_{u}\Bigl[Vf(x,u)-u+W_{t+1}\bigl(x+f(x,u)\bigr)\Bigr].\]
查看完整验算程序
# Chapter 3. p31: Bellman's backward induction finds the same pulse.
# State x on a grid of 1001 points in [0, 1] with linear interpolation;
# control u in 11 levels from 0 to 5000 yuan per day, 500 apart; value
# W_t(x) = max_u [ V f(x, u) - u + W_{t+1}(x + f(x, u)) ], W_60 = 0.
# Nothing assumes bang-bang: the grid is free to choose middle levels.
# The policy is also read as a feedback rule: on day t spend while the
# reached share x is below a threshold x*_t. Parameters as in p12.
import math
M, MARGIN, P, Q, BETA, UMAX, T = 20000, 30.0, 0.001, 0.08, 4e-6, 5000.0, 60
V = MARGIN * M
N = 1000
LEVELS = [UMAX * k / 10 for k in range(11)]
def f(x, u):
return (P + BETA * u + Q * x) * (1 - x)
def interp(w, x):
if x >= 1.0:
return w[N]
i = int(x * N)
a = x * N - i
return w[i] * (1 - a) + w[i + 1] * a
w = [0.0] * (N + 1)
policy = [None] * T
for t in range(T - 1, -1, -1):
nw, pol = [0.0] * (N + 1), [0.0] * (N + 1)
for i in range(N + 1):
x = i / N
best, bu = None, 0.0
for u in LEVELS:
dx = f(x, u)
val = V * dx - u + interp(w, x + dx)
if best is None or val > best:
best, bu = val, u
nw[i], pol[i] = best, bu
w, policy[t] = nw, pol
x, us, cash = 0.0, [], []
for t in range(T):
u = policy[t][int(round(x * N))]
dx = f(x, u)
us.append(u)
cash.append(V * dx - u)
x += dx
print("DP value at x=0: %.0f simulated path profit: %.0f"
% (w[0], math.fsum(cash)))
print("spend path days 0-9: " + " ".join("%d" % u for u in us[:10]))
mid = [u for u in us if 0 < u < UMAX]
print("days at the cap: %d, days at a middle level: %d, total %.0f"
% (len([u for u in us if u == UMAX]), len(mid), math.fsum(us)))
print("day threshold x*_t (spend the cap while x is below it)")
for t in (0, 2, 4, 6, 8, 10, 20, 40):
xs = [i / N for i in range(N + 1) if policy[t][i] > 0]
print("%3d %s" % (t, "%.3f" % max(xs) if xs else "none"))唯一数据源与引用
范雅福(Fan, Y.). (2026). 《庞特里亚金极大值原理、新品冷启动广告脉冲烧钱速率与口碑接棒最优控制母典》第 3 章「从最后一天往回推:贝尔曼逆推、网格检查与硬预算」(版本 1.0.0). 5D Library. DOI 待发布
DOI 待发布 · 专著页正式发布后上线(本站暂无专著页) · 5DL-B-241 · v1.0.0 · CC-BY-NC-ND-4.0
BibTeX
@book{fan2026241,
author = {Fan, Yafu},
title = {Pontryagin's Maximum Principle at Product Launch: Optimal Ad-Pulse Burn Rates and the Hand-Over to Word of Mouth},
year = {2026},
publisher = {5D Library},
edition = {1.0.0},
note = {DOI pending}
}